obscura/minimind
1
0
Fork 0
mirror of https://github.com/jingyaogong/minimind.git synced 2026-10-13 04:27:24 +00:00
Code Issues Packages Projects Releases Wiki Activity

[docs] 在客观评测表下补充各数据集的测量分辨率

Browse Source 
lm_eval 对每个指标都会输出 stderr，但表格只保留了点估计。这里按测试集
大小给出各任务的 ± 标准误（obqa ±1.9 到 cmmlu ±0.4），并说明比较两个
模型时差值的标准误约为单项的 √2 倍，方便读者判断表中哪些差距是可以
解读的。中英文 README 同步。
This commit is contained in:
Justin Guese
2026-09-03 17:10:52 +07:00
parent 7a6fddd63a
commit 77946fda00
2 changed files with 18 additions and 0 deletions
Download Patch File Download Diff File Expand all files Collapse all files
README.md
+9
View File
@@ -1599,6 +1599,15 @@ MiniMind 的数据规模远小于表中其他模型，且训练比例偏向中
| [SmolLM2-135M](https://huggingface.co/HuggingFaceTB/SmolLM2-135M-Instruct) | HuggingFace | 135M | 24.44 / 24.71 | 58.50 / 68.17 / 32.80 / 43.15 / 39.46 |
| [Aquila-135M](https://huggingface.co/BAAI/Aquila-135M-Instruct) | BAAI | 135M | 25.19 / 25.10 | 54.59 / 67.52 / 34.40 / 41.67 / 39.66 |
> **这张表的分辨率**：`lm_eval` 对每个指标都会同时输出一个 `stderr`，它主要由测试集大小决定，因此下面的数值对表中所有模型都近似适用（随准确率略有浮动）：
>
> | 数据集 | obqa | ceval | piqa | siqa | arc | hellaswag | cmmlu |
> |---|---|---|---|---|---|---|---|
> | 样本量 | 500 | ~1.3k | 1838 | 1954 | 2376 | 10042 | 11582 |
> | ± 标准误 | **1.9** | 1.2 | 1.2 | 1.1 | 0.9 | 0.5 | **0.4** |
>
> 比较两个模型时，差值的标准误约为单项的 `√2` 倍。因此 obqa 上小于约 5 个百分点、ceval / piqa / siqa 上小于约 3 个百分点的差异，都不足以区分两个模型——这也解释了为什么本表中同一量级的模型排序不宜过度解读。
<details>
<summary><strong>补充说明（来源/无污染/复现）</strong></summary>
README_en.md
+9
View File
@@ -1597,6 +1597,15 @@ MiniMind is trained on far less data than the other models listed here, and its
| [SmolLM2-135M](https://huggingface.co/HuggingFaceTB/SmolLM2-135M-Instruct) | HuggingFace | 135M | 24.44 / 24.71 | 58.50 / 68.17 / 32.80 / 43.15 / 39.46 |
| [Aquila-135M](https://huggingface.co/BAAI/Aquila-135M-Instruct) | BAAI | 135M | 25.19 / 25.10 | 54.59 / 67.52 / 34.40 / 41.67 / 39.66 |
> **The resolution of this table**: `lm_eval` reports a `stderr` alongside every metric. It is set mainly by the size of the test set, so the values below apply approximately to every model in the table (they shift slightly with accuracy):
>
> | dataset | obqa | ceval | piqa | siqa | arc | hellaswag | cmmlu |
> |---|---|---|---|---|---|---|---|
> | samples | 500 | ~1.3k | 1838 | 1954 | 2376 | 10042 | 11582 |
> | ± stderr | **1.9** | 1.2 | 1.2 | 1.1 | 0.9 | 0.5 | **0.4** |
>
> When comparing two models, the standard error of the difference is about `√2` times the single-model value. So a gap smaller than roughly 5 points on obqa, or roughly 3 points on ceval / piqa / siqa, does not separate two models — which is why the ordering of similarly-sized models in this table should not be over-read.
<details>
<summary><strong>Additional note (source / no contamination / reproduction)</strong></summary>